A 1D CNN for high accuracy classification and transfer learning in motor imagery EEG-based brain-computer interface.

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Brain-computer interfaces (BCI) aim to create communication pathways between brain activity and external devices.
• Electroencephalography (EEG)-based BCIs, particularly those using motor imagery (MI), show significant potential for applications.
• Developing BCI techniques that minimize extensive data pre-processing is crucial for practical implementation.

Purpose of the Study:

• To propose a novel 10-layer 1D-CNN approach for classifying brain states from EEG signals.
• To develop a transfer learning method for efficient individual-specific BCI model customization.
• To enhance BCI development by utilizing raw EEG signals with minimal pre-processing.

Main Methods:

• Implementation of a 10-layer 1D-CNN model for classifying five brain states (four MI classes and a baseline).
• Utilization of a data augmentation algorithm and a limited number of EEG channels.
• Development of a transfer learning strategy for feature extraction from group data and fine-tuning for individual users with 12-minute data.

Main Results:

• The proposed 1D-CNN model achieved a 99.38% accuracy at the group level on the 'EEG Motor Movement/Imagery Dataset', surpassing current state-of-the-art.
• The transfer learning approach demonstrated an average accuracy of 99.46% for individual-specific BCI models.
• The methods effectively classify brain states from raw EEG signals, reducing the need for labor-intensive pre-processing.

Conclusions:

• The developed 1D-CNN and transfer learning methods offer a highly accurate and efficient approach for BCI applications.
• These techniques can facilitate the creation of future BCI systems using portable, few-channel recording devices.
• The individual-based training approach accelerates BCI system adaptation, paving the way for more personalized neurotechnology.